| Home | Organizers | Proceedings Editors | Proceedings Contributors | Search | | ||
|
|
|||
|
||||
|
|
||||
| Title: | EXPERIMENTAL STUDIES ON REPRESENTATION COMPLEXITY AND ERROR RATES OF ITERATIVELY COMPOSED FEATURES | |
| DOI No: | 10.1142/9781860948534_0015 | |
| Source: | INNOVATIVE APPLICATIONS OF INFORMATION TECHNOLOGY FOR THE DEVELOPING WORLD (pp 92-96) | |
| Author(s): | K. HARAGUCHI
Department of Applied Mathematics and Physics, Graduate School of Informatics, Kyoto University, Kyoto, Japan H. NAGAMOCHI Department of Applied Mathematics and Physics, Graduate School of Informatics, Kyoto University, Kyoto, Japan T. IBARAKI Department of Informatics, School of Science and Technology, Kwansei Gakuin University, Sanda, Japan |
|
| Abstract: | We consider constructing a classifier c : {0, 1}n ↦ {0,1} from a training set of n-dimensional binary vectors. A classifier is usually represented by a representation model such as decision tree (DT) and iteratively composed feature (ICF). In this paper, we discuss the following issues through computational experiments: 1) relation between the representation complexity γ, the error on the training set e and the expected error E of an ICF classifier and 2) comparison between ICF and DT. For 1), we observe that a simpler ICF is more likely to attain a small E, among the ICFs having the same e. For 2), we see that DTs generated by C4.5 and ICFs generated by our algorithm are competitive in terms of E. | |
| Full Text: | View full text in PDF format (249KB) | |
| TOC: | Back to Table of Contents | |
|
||